FET-input op-amps behave differently from bipolar-input op-amps. Take a look at Figure 4.9, taken from a TL072 working in shunt and in series configuration with a 5 Vrms output.
Figure 4.9: A TL072 shunt-feedback stage using 10 and 22 kΩ resistors shows low distortion. The series version is much worse due to the impedance of the NFB network, but it can be made the same as the shunt case by adding cancellation source resistance in the input path. No external loading, test level 5 Vrms, supply ±18 V
The circuits are as in Figure 4.3(a) and (b), except that the resistor values have to be scaled up to 10 and 22 kΩ because the TL072 is nothing like so good at driving loads as the 5532. This unfortunately means that the inverting input is seeing a source resistance of 10k||22k = 6.9k, which introduces a lot of CM distortion in the series case – five times as much at 20 kHz as for the shunt case. Adding a similar resistance in the input path cancels out this distortion, and the trace then is the same as the 'Shunt' trace in Figure 4.9. Disconcertingly, the value that achieved this was not 6.9k, but 9k1. That means adding -113 dBu of Johnson noise, so it's not always appropriate.
It's worth mentioning that the flat part of the 'Shunt' trace below 10 kHz is not noise, as it would be for the 5532; it is distortion.
A voltage-follower has no inconvenient medium-impedance feedback network, but it does have a much larger CM voltage. Figure 4.10 shows a voltage-follower working at 5 Vrms. With no source resistance the distortion is quite low, due to the 100% NFB, but as soon as a 10 kΩ source resistance is added we are looking at 0.015% at 10 kHz.
Figure 4.10: A TL072 voltage-follower working at 5 Vrms with a low source resistance produces little distortion (RS = 0R), but adding a 10 kΩ source resistance makes things much worse (RS = 10k). Putting a 10 kΩ resistance in the feedback path as well gives complete cancellation of this extra distortion (Cancel). Supply ±18 V
Once again, this can be cured by inserting an equal resistance in the feedback path of the voltage-follower, as in Figure 4.3(d) above. This gives the 'Cancel' trace in Figure 4.10. Adding resistances for distortion cancellation in this way has the obvious disadvantage that they introduce extra Johnson noise into the circuit.
Another point is that stages of this kind are often driven from pot wipers, so the source impedance is variable, ranging between zero and one-quarter of the pot track resistance. Setting a balancing impedance in the other op-amp input to a mid-value, i.e. one-eighth of the track resistance, should reduce the average amount of input distortion, but it is inevitably a compromise.
With JFET inputs the problem is not the operating currents of the input devices themselves, which are negligible, but the currents drawn by the non-linear junction capacitances inherent in field-effect devices. These capacitances are effectively connected to one of the supply rails. For P-channel JFETs, as used in the input stages of most JFET op-amps, the important capacitances are between the input JFETs and the substrate, which is normally connected to the V- rail (see Jung ).
According to the Burr-Brown data sheet for the OPA2134, 'The P-channel JFETs in the input stage exhibit a varying input capacitance with applied CM voltage.' It goes on to recommend that the input impedances should be matched if they are above 2 kΩ.
Common-mode distortion can be minimized by running the op-amp off the highest supply rails permitted, though the differences are not large. In one test on a TL072, going from ±15 to ±18 V rails reduced the distortion from 0.0045% to 0.0035% at 10 kHz.